CAREER: Flexible and Robust Reasoning in Natural Language

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。随着大型神经网络模型的发展，嵌入搜索引擎和数字助理中的现代问答系统得到了极大的改进。当用户问一个简单的问题时，这些系统通常可以直接返回答案，而不仅仅是链接到网页。然而，这些系统在处理更复杂的问题时仍然会失败，当它们失败时，它们可能会误导用户。它们缺乏人类拥有的一种重要能力：对所看到的信息进行推理和综合、检索和整合附加信息并得出合理结论的能力。这个CAREER项目旨在通过开发“思考”文本证据的系统来解决这一缺点，从而获得可以向用户解释的更可靠的答案。这些进步符合构建可信赖的人工智能系统的更广泛思路，这些系统可以明确显示其工作，并在部署之前和部署期间进行审计。该项目通过开发一个以自然语言推理的基于学习的系统，专门解决了问题回答和事实检查的问题。该系统将文本作为输入，然后应用预训练的神经网络模型来重新表述文本，从中得出结论，并最终检查索赔或验证答案。这个过程产生了一系列逻辑相连的语句，人类可以理解。这一结果由两个模块实现。首先，演绎模块重复组合两个语句，并根据输入生成第三个语句，封装公共逻辑规则。其次，验证者确定最终推断的证据是否证实了最初的主张。这两个系统都是由像T5这样的预训练模型构建的，这些模型已经证明了强大的泛化能力。为这些模型收集训练数据是一个核心挑战；该项目的方法混合了多种策略，包括合成数据生成和人在循环中的注释。这些技术应用于问题回答和事实检查的领域，在这些问题中，提供额外的解释和证明，而不仅仅是给出最大努力的答案，对于创建可用的系统至关重要。这个系统为NLP工具知道他们不知道的东西铺平了道路，为最终用户提供了可解释性，并使系统开发人员能够更好地理解和改进他们的模型。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Modern question answering systems, embedded in search engines and digital assistants, have improved dramatically with the development of large neural network models. When a user asks a simple question, these systems can typically return an answer directly rather than just linking to a webpage. However, these systems still fail on more complex questions, and when they fail, they may mislead their users. They lack an important capability that humans have: the ability to reason about and synthesize the information they see, retrieve and integrate additional information, and arrive at a justified conclusion. This CAREER project aims to address this shortcoming by developing systems that "think through" textual evidence, leading to more reliable answers that can be explained to a user. Such advances fit into a broader thread of building trustable AI systems that explicitly show their work and are auditable before and during their deployment.This project specifically addresses the problems of question answering and fact-checking by developing a learning-based system that reasons in natural language. The system takes text as input, then applies pre-trained neural network models to reformulate that text, derive conclusions from it, and eventually check a claim or verify an answer. This process produces a series of logically connected statements understandable by a human. This outcome is enabled by two modules. First, a deduction module repeatedly combines two statements and generates a third that follows from the inputs, encapsulating common logical rules. Second, a verifier determines whether the final deduced evidence validates the original claim. Both systems are built from pre-trained models like T5 that have demonstrated strong generalization capabilities. Collecting training data for these models constitutes a core challenge; the project's approach blends multiple strategies including synthetic data generation and human-in-the-loop annotation. These techniques are applied to the domains of question answering and fact checking, problems where providing additional explanation and justification instead of just giving a best-effort answer are essential to make usable systems. This system paves the way for NLP tools that know what they don't know, provide interpretability for end users, and enable system developers to better understand and improve their models.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.